Re: [LKP] [sched/fair] 6c8116c914: stress-ng.mmapfork.ops_per_sec -38.0% regression

From: Vincent Guittot
Date: Mon Apr 27 2020 - 08:46:25 EST


On Mon, 27 Apr 2020 at 13:35, Hillf Danton <hdanton@xxxxxxxx> wrote:
>
>
> On Mon, 27 Apr 2020 11:03:58 +0200 Vincent Guittot wrote:
> > On Sun, 26 Apr 2020 at 14:42, Hillf Danton wrote:
> > >
> > > On 4/21/2020 8:47 AM, kernel test robot wrote:
> > > >
> > > > Greeting,
> > > >
> > > > FYI, we noticed a 56.4% improvement of stress-ng.fifo.ops_per_sec due to commit:
> > > >
> > > >
> > > > commit: 6c8116c914b65be5e4d6f66d69c8142eb0648c22 ("sched/fair: Fix condition of avg_load calculation")
> > > > https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git master
> > > >
> > > > in testcase: stress-ng
> > > > on test machine: 88 threads Intel(R) Xeon(R) CPU E5-2699 v4 @ 2.20GHz with 128G memory
> > > > with following parameters:
> > > >
> > > > nr_threads: 100%
> > > > disk: 1HDD
> > > > testtime: 1s
> > > > class: scheduler
> > > > cpufreq_governor: performance
> > > > ucode: 0xb000038
> > > > sc_pid_max: 4194304
> > > >
> > >
> > > We need to handle group_fully_busy in a different way from
> > > group_overloaded as task push does not help grow load balance
> > > in the former case.
> >
> > Have you tested this patch for the UC above ? Do you have figures ?
> >
> No I am looking for a box of 88 threads. Likely to get access to it in
> as early as three weeks.
>
> > > --- a/kernel/sched/fair.c
> > > +++ b/kernel/sched/fair.c
> > > @@ -8744,30 +8744,20 @@ find_idlest_group(struct sched_domain *s
> > >
> > > switch (local_sgs.group_type) {
> > > case group_overloaded:
> > > - case group_fully_busy:
> > > - /*
> > > - * When comparing groups across NUMA domains, it's possible for
> > > - * the local domain to be very lightly loaded relative to the
> > > - * remote domains but "imbalance" skews the comparison making
> > > - * remote CPUs look much more favourable. When considering
> > > - * cross-domain, add imbalance to the load on the remote node
> > > - * and consider staying local.
> > > - */
> > > -
> > > - if ((sd->flags & SD_NUMA) &&
> > > - ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
> > > + if (100 * local_sgs.avg_load <= sd->imbalance_pct * (idlest_sgs.avg_load + imbalance))
> > > + return idlest;
> >
> > So you have completely removed the NUMA special case without explaining why.
> >
> That was for the local domain that is lightly loaded, as the comment says,
> it now is overloaded.

The load value is not linked to the overloaded state of the group as
you can be overloaded but still have a low load especially with cgroup

That's also why there are 2 type of comparison:
an absolute comparison for low load value
and a proportional comparison for normal/high value

>
> > And you have also removed the tests for small load.
> >
> It is a heuristic I want to avoid. It can be replaced with the load of the
> task in question as best effort.
>
> > Could you explain the rationale behind all these changes ?
> >
> > Also keep in mind that the current version provide +58% improvement
> > for stress-ng.fifo
> >
> Yes it's great. I'm on the minor one.
>
> > > + if (local_sgs.avg_load > idlest_sgs.avg_load + imbalance)
> > > + return idlest;
> > > + else
> > > return NULL;
> > >
> > > + case group_fully_busy:
> > > /*
> > > - * If the local group is less loaded than the selected
> > > - * idlest group don't try and push any tasks.
> > > + * Pushing task to the idlest group will make the target group
> > > + * overloaded, leaving the local group that is overloaded fully busy,
> > > + * thus we earn nothing except for the exchange of group types.
> >
> > For this case both local and idlest are fully busy and in this case
> > one will become overloaded so you must compare the load to be fair in
> > the spread of load
> >
> It may be sooner than thought that the newly overloaded group is looking to
> push task out, and we'll see a task ping-pong if it happens.
>
>
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -8683,15 +8683,11 @@ find_idlest_group(struct sched_domain *s
> struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
> struct sg_lb_stats local_sgs, tmp_sgs;
> struct sg_lb_stats *sgs;
> - unsigned long imbalance;
> struct sg_lb_stats idlest_sgs = {
> .avg_load = UINT_MAX,
> .group_type = group_overloaded,
> };
>
> - imbalance = scale_load_down(NICE_0_LOAD) *
> - (sd->imbalance_pct-100) / 100;
> -
> do {
> int local_group;
>
> @@ -8743,31 +8739,26 @@ find_idlest_group(struct sched_domain *s
> return idlest;
>
> switch (local_sgs.group_type) {
> - case group_overloaded:
> case group_fully_busy:
> - /*
> - * When comparing groups across NUMA domains, it's possible for
> - * the local domain to be very lightly loaded relative to the
> - * remote domains but "imbalance" skews the comparison making
> - * remote CPUs look much more favourable. When considering
> - * cross-domain, add imbalance to the load on the remote node
> - * and consider staying local.
> - */
> -
> - if ((sd->flags & SD_NUMA) &&
> - ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
> - return NULL;
> -
> - /*
> - * If the local group is less loaded than the selected
> - * idlest group don't try and push any tasks.
> - */
> - if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
> - return NULL;
> -
> - if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
> - return NULL;
> - break;
> + return NULL;
> + case group_overloaded:
> + /* No push if balanced */
> + if (100 * local_sgs.avg_load > sd->imbalance_pct *
> + idlest_sgs.avg_load) {
> + unsigned long avg_load;
> +
> + avg_load = task_h_load(p) + idlest_sgs.group_load;
> + avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
> + idlest_sgs.group_capacity;
> +
> + if (100 * local_sgs.avg_load <= sd->imbalance_pct *
> + avg_load)
> + return idlest;
> +
> + if (local_sgs.avg_load > avg_load)
> + return idlest;
> + }
> + return NULL;
>
> case group_imbalanced:
> case group_asym_packing:
>